Skip to content
AI.info

The Pulse

RadixArk releases DaisyRose-27B with post-training efficiency claims

DaisyRose-27B is a reinforcement-learning post-train of Qwen3.8-27B, with checkpoints available under Apache 2.0. RadixArk reports higher benchmark scores and lower GPU time than a comparable open-source training setup.

RadixArk releases DaisyRose-27B with post-training efficiency claims

DaisyRose-27B raises the Terminal-Bench 4.1 pass@1 score from 5.6% for its base model to 12.0%, according to its developer RadixArk. The model, released on November 18, is RadixArk’s first and builds on Qwen3.8-27B using reinforcement learning.

The company says it trained the model on about 1,200 curated agent tasks over 18 days using 144 H200 GPUs. Sessions in training ran up to 1 million tokens, with summarization used to span multiple trajectories.

Model and benchmark results

The release retains the base model’s architecture, tokenizer, chat template, image understanding and 256K-token context window. Checkpoints are available in BF16, FP8 and NVFP4 formats under the Apache 2.0 license.

The model reasons before answering by default, returning its reasoning separately from its final answer. The company recommends serving the model with SGLang and DSpark speculative decoding.

RadixArk reports gains across all seven benchmarks in its comparison with Qwen3.8-27B. On DeepSWE v1.1, the reported score rose from 42.2 to 46.2; on IFBench, it increased from 79.5 to 82.4. For each benchmark, the reported score is the higher of the BF16 and FP8 runs; where both were evaluated, they differed by less than one point.

Tasks built around mistakes

RadixArk and Eigent AI developed thousands of candidate tasks across areas including machine learning, software engineering and science. About 1,200 entered the final training set.

RadixArk ran the base model on each task and redesigned tasks around mistakes it wanted to address, adjusting difficulty to produce a useful solve rate. The tasks required multiple dependent steps, and attempts ran in sandboxes where the agent could execute commands, edit files and be graded.

Each attempt could run for up to 1.5 hours, with more than 100 tool calls and more than 256K output tokens. Tasks that could be passed without solving the intended problem, or that had design flaws, were fixed or removed after review.

Training setup and cost claims

The release was trained with Miles, RadixArk’s open-source reinforcement-learning framework, in an asynchronous setup: agents continued working while the model learned. In the setup, 128 H200 GPUs ran inference on agent tasks and another 16 ran the trainer; updated models were sent back to the inference GPUs after each step.

RadixArk says its internal backend put 37% of each GPU’s peak compute into useful training work, compared with 16% for an open-source trainer built on Megatron in comparable runs. On that basis, the company says the same training needs 57% less GPU time.

Counting the time agents spent on tasks, the release run trained at about 16,800 tokens per second across its 144 GPUs, compared with about 8,300 on the same GPUs for the open-source trainer, according to RadixArk.

Sandbox infrastructure

The sandbox service is built on the open-source AgentENV project. It ran about 944,000 sandboxes for the project between mid-September and mid-November, with up to 1,566 running concurrently, RadixArk says.

Across nine weeks, 99% of sandbox creations succeeded, and that figure has exceeded 99.95% since late October, according to the company. The service grew from one machine to as many as 12 and received ten upgrades, nearly all while training was underway.

The company estimates that cloud infrastructure for its sandboxes cost about 80% less than running the same sandboxes with popular cloud sandbox providers. Its service lets each sandbox’s memory allocation expand or shrink with usage, allowing more sandboxes to share a machine.

Source

  1. RadixArk

Related stories

Explore

More articles